PoBO: A Polynomial Bounding Method for Chance-Constrained Yield-Aware Optimization of Photonic ICs

Zichang He, Zheng Zhang

TL;DR

PoBO (Polynomial Bounding Optimization) is a new framework for designing photonic integrated circuits (ICs) that maximizes performance while strictly guaranteeing a specific manufacturing yield. By replacing traditional, overly-conservative "moment-based" statistical bounds with a tighter polynomial bounding method, PoBO discovers high-performance designs that previous methods would have deemed "too risky."

What Problem Does This Paper Address?

Manufacturing a photonic IC is a game of microscopic precision. Tiny variations in geometry (process variations) can lead to significant performance drops or total "yield loss."

Traditionally, engineers have tried to maximize yield—essentially trying to make as many working chips as possible. However, this often leads to "safe" but mediocre designs. A more recent approach is chance-constrained optimization, where you maximize a performance metric (like bandwidth) while ensuring the probability of failure stays below a certain threshold (e.g., 1%).

The mathematical challenge is that these "chance constraints" are notoriously difficult to calculate. Previous solutions used "moment-bounding" (like the Chebyshev-Cantelli inequality), which only looks at the mean and variance. Because these bounds are very loose, they are "over-conservative," shrinking the pool of feasible designs and often missing the optimal solution entirely.

Methods

The authors introduce PoBO, which uses a sequence of three main steps to solve the optimization problem more accurately:

  1. Surrogate Modeling: Instead of running thousands of expensive circuit simulations, they build a "surrogate" (a mathematical stand-in) using polynomial chaos expansions. This model predicts how the circuit will behave under different variations.
  2. Optimal Polynomial Kinship Functions: This is the paper's core innovation. A "kinship function" is a mathematical tool used to approximate the "pass/fail" step function of yield. The authors developed a way to find the best polynomial to bound this function. Unlike previous methods, they relaxed the requirements for "convexity" and "independence," allowing the bound to be much tighter to the actual failure probability.
  3. Global Polynomial Optimization: Because the objective and the bounds are now expressed as polynomials, the authors can use a global solver (like GloptiPoly) to find the absolute best design variables, rather than getting stuck in "local optima."

Datasets

The method was evaluated using three distinct benchmarks:

  • Synthetic Function: A mathematical test case designed to simulate non-Gaussian, correlated variations where traditional methods typically struggle.
  • Mach-Zehnder Interferometer (MZI): A third-order MZI benchmark where the goal was to maximize bandwidth while meeting constraints on crosstalk and attenuation.
  • Microring Add-Drop Filter: A benchmark involving three silicon microrings, aimed at maximizing bandwidth while maintaining a specific extinction ratio and passband roughness.

Key Insights and Ideas

  • Tighter Bounds = Better Performance: By using higher-order polynomials instead of just first and second moments (mean/variance), the "feasible region" for the design expands. This allows the optimizer to explore high-performance areas that moment-based methods would have blocked off as "unsafe."
  • Relaxing Constraints for Accuracy: Most existing kinship functions are required to be convex for easier math. By dropping the convexity requirement, the authors created a bounding function that much more closely hugs the actual failure boundary.
  • Simulation Efficiency: PoBO reuses the "quadrature points" (specific simulation samples) from the surrogate modeling step to calculate the yield risk. This means it provides a rigorous statistical bound without requiring any extra circuit simulations beyond what was needed to build the model.

Key Results

  • Superior Performance: In the MZI benchmark, PoBO achieved a bandwidth of 193.28 GHz compared to 189.35 GHz for the moment-bounding method at the same risk level.
  • Feasibility where others fail: In high-yield scenarios (e.g., 99% yield requirement), the moment-bounding method often failed to find any valid design because its bounds were too conservative. PoBO successfully found optimal designs in these cases.
  • Massive Speedup vs. Bayesian Methods: Compared to Bayesian Yield Optimization (BYO), PoBO required 30× to 58× fewer circuit simulations to reach a solution, making it far more practical for complex IC designs.

Limitations and Future Work

  • The Curse of Dimensionality: While PoBO handles many process variations well, the underlying polynomial optimization solvers (SDP relaxations) can become computationally expensive as the number of design variables increases.
  • Joint Constraints: The current work focuses on individual chance constraints. The authors note that future work will involve developing "multivariate kinship functions" to handle multiple performance constraints simultaneously more efficiently.
  • Beyond Polynomials: While the current framework relies on polynomial surrogates, the authors suggest the bounding logic could potentially be extended to other types of performance models in the future.

Summary updated

Abstract

Conventional yield optimization algorithms try to maximize the success rate of a circuit under process variations. These methods often obtain a high yield but reach a design performance that is far from the optimal value. This article investigates an alternative yield-aware optimization for photonic ICs: we will optimize the circuit design performance while ensuring a high yield requirement. This problem was recently formulated as a chance-constrained optimization, and the chance constraint was converted to a stronger constraint with statistical moments. Such a conversion reduces the feasible set and sometimes leads to an over-conservative design. To address this fundamental challenge, this article proposes a carefully designed polynomial function, called optimal polynomial kinship function, to bound the chance constraint more accurately. We modify existing kinship functions via relaxing the independence and convexity requirements, which fits our more general uncertainty modeling and tightens the bounding functions. The proposed method enables a global optimum search for the design variables via polynomial optimization. We validate this method with a synthetic function and two photonic IC design benchmarks, showing that our method can obtain better design performance while meeting a prespecified yield requirement. Many other advanced problems of yield-aware optimization and more general safety-critical design/control can be solved based on this work in the future.
BibTeX
@article{he2022pobo,
  author = {Zichang He and Zheng Zhang},
  title = {{PoBO: A Polynomial Bounding Method for Chance-Constrained Yield-Aware Optimization of Photonic ICs}},
  journal = {IEEE Trans. Comput. Aided Des. Integr. Circuits Syst.},
  year = {2022},
  url = {https://openreview.net/forum?id=wCVVOE5VuBa},
}

Loading related content...

Figure